Curious builder0 XP earned · 300 to level 2

0 daysFinish a lesson to begin

Badge collection0 of 6 unlocked

51 small wins to finish your pathNext question →

Q22IntermediateSystem design

Design AI-powered semantic product search for an e-commerce site with 10M products.

30-second answerSay your answer out loud first, then reveal.

Combine lexical search (BM25 over titles and attributes) with dense retrieval (product embeddings) for recall. Use an LLM or small model for query understanding: spelling correction, intent, extracting filters like price, colour, brand, transliteration. Then a learning-to-rank model that blends relevance with business signals (CTR, conversions, stock, margin). Keep LLM calls off the hot path where possible (cache query understanding, precompute product enrichments), since search needs about 100–200ms latency.

Product search architecture: offline, the catalog is enriched by an LLM, embedded and indexed; online, a query goes through cached or small-model query understanding, then BM25 and vector search with filters, candidates (~500) are merged and ranked by learning-to-rank on relevance, CTR, conversions, stock and personalisation, returning top results in ~50-150ms.

Design points

Latency budget: total ~150ms. A frontier LLM call per query is too slow and too costly at high QPS, so use:
• a cache of query-understanding results for frequent queries (the head of the distribution repeats a lot),
• small fine-tuned models for filter extraction on tail queries.
Offline LLM enrichment: fill missing attributes, normalise colours and sizes, generate searchable descriptions, translate. Batch processing is cheap at catalogue scale.
Multilingual / transliterated queries ("joote", "saree for shaadi"): query normalisation plus multilingual embeddings.
Ranking: a learning-to-rank model (e.g. gradient-boosted) trained on clicks and purchases, with position-bias correction.
Evaluation: offline nDCG on judged query–product pairs (LLM-assisted relevance labelling + human audit), then online A/B tests on CTR, add-to-cart, conversion, null-result rate.

Optional: a conversational shopping assistant layered on top for exploratory queries ("gift for a 10-year-old who likes science").

PreviousDesign a natural-language analytics assistant (text-to-SQL) over a company data warehouse.

You understood something today that you didn't yesterday.

Design AI-powered semantic product search for an e-commerce site with 10M products.

Design points

Related